A Deep Path Planning Algorithm Based on CNNs for Perception Images

Path planning for robots navigation, commercial computer games, off-line map applications and many other fields is an ongoing research. There have raised several methods derived from the traditional A-star algorithm due to its efficiency in the past few years. As for the limitations of algorithms in global path planning, we introduce a novel method based on deep learning in this paper. We present a novel path planning algorithm combined with convolutional neural networks (CNNs) to learn a target-oriented end-to-end model from the input of images. The deep neural network proved to be efficient and effective in feature extracting in our experiments too. The model can transfer the scene understanding and navigation knowledge gained from one environment to another unseen ones. Finally, this method can not only maintain the optimality of the path, but can also greatly accelerate the computation.